BDRI: block decomposition based on relational interaction for knowledge graph completion

• Published in: Data mining and knowledge discovery Vol. 37; no. 2; pp. 767 - 787
• Main Authors: Yu, Mei, Guo, Jiujiang, Yu, Jian, Xu, Tianyi, Zhao, Mankun, Liu, Hongwei, Li, Xuewei, Yu, Ruiguo
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2023; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Chemistry and Earth Sciences; Computer Science; Data Mining and Knowledge Discovery; Decomposition; Graphical representations; Information Storage and Retrieval; Knowledge representation; Mathematical analysis; Physics; Statistics for Engineering; Tensors; Tensor decomposition; Inverse relation; Knowledge graph completion
• DOI: 10.1007/s10618-023-00918-8
• ISSN: 1384-5810

Knowledge graphs (KGs) are large-scale semantic networks designed to describe real-world facts. However, existing KGs typically contain only a small subset of all possible facts. Knowledge graph completion (KGC) is a task of inferring missing facts based on existing facts, which can help KGs become more complete. Tensor decomposition algorithms have proved promising for KGC problems. In this paper, we propose block decomposition based on relational interaction for knowledge graph completion (BDRI), a novel and robust model based on block term decomposition of the binary tensor representation of knowledge graph triples. Further, BDRI considers that the inverse relation, as one of the most important relation types, not only occupies a large proportion in real-world facts but also has an impact on other relation types. Although some existing models also take into account the importance of inverse relations, it is not enough to learn inverse relations independently. BDRI strengthens the fusion of forward relations and inverse relations by introducing inverse relations into the model in an enhanced way. We prove BDRI is full expressiveness and derive the bound on its entity and relation embedding dimensionality and smaller than the bound of SimplE and ComplEx. Experimental results on five public datasets show the effectiveness of BDRI.